Re: [Kaldi-developers] queries on Kaldi

[Vassil P. <vas...@gm...>]

Hi Robert,

I think the VoxForge recipe is a very good idea. I am not sure what
language model should be used for the decoding part of the recipe, but
since it will be just a demo system with free data, it doesn't matter
that much and probably just training a "cheating" LM on all
transcripts is OK.

I agree with Arnab, that it would be the best if you base your recipe
on wsj/s5. In fact I think it shouldn't be very difficult to modify it
to use the VoxForge data, because for example in egs/rm/s5 just the
data normalization scripts("local/") are specific to RM, and steps/
and utils/ directories are just symlinks to wsj/s5.

Vassil

On Fri, Jun 15, 2012 at 2:21 PM, Arnab Ghoshal <ar...@gm...> wrote:
> On Mon, Jun 11, 2012 at 3:09 PM, Robert Mullins <r_p...@ya...> wrote:
>> I downloaded and installed the Kaldi code on my machine, however I do not
>> have the data from the LDC RM discs. I was just wondering whether anyone has
>> used the voxforge speech corpus for training your system? And if not is
>> there a reason and would it be a worthwhile exercise?
>
> Hi Robert,
>
> you could use the egs/rm/s4 recipe that uses a free subset of RM. Here
> is a blog post about it:
> http://vpanayotov.blogspot.co.uk/2012/02/poor-mans-kaldi-recipe-setup.html
>
> There may be people on this list who have used voxforge. You are
> certainly welcome to contribute a recipe for it.
>
>> I would also be interested in seeing an example of how to use the kaldi
>> system as an end to end speech recognition system
>
> That's exactly what Kaldi is meant to be. Once you go through the
> steps of rm/s4 you will get a hang of it. If you want to start
> building a voxforge recipe, I will recommend that you start from
> egs/wsj/s5.
>
> -Arnab